AI in GxP Manufacturing

The life sciences sector is not moving slowly on AI. On the research and development side, teams are already using machine learning tools for cell pattern recognition, regulatory submissions, and clinical data analysis. The regulatory agencies have shown a notable ability to engage with these applications and keep pace with how they are evolving. That side of the business is moving with real momentum.

On the manufacturing and compliance side, the picture is more nuanced. And according to Bill Burke, CEO of Merit Solutions, a software company exclusively serving life science and regulated industry manufacturers with GxP-compliant ERP and quality management software built on Microsoft Dynamics 365, and Matt Birtwistle, Manufacturing Industry Lead at Nexer, the organisations that get ahead of the governance question now will be in a considerably stronger position than those who encounter it reactively.

AI Agents in GMP Environments

Microsoft and other platform vendors are building AI agents directly into their software. For most industries, this is straightforwardly positive. For GMP-regulated manufacturers, it warrants careful attention.

Bill Burke raised a specific concern in their conversation: as Dynamics 365 and other platforms continue to expand their agent capabilities, manufacturers who do not have a clear view of which agents are active and what processes they touch could inadvertently introduce AI-driven decisioning into GMP workflows that require documented human oversight. The challenge is not malicious. It is simply that agents designed to add efficiency for a general manufacturer may intersect with processes that a regulated manufacturer has classified as significant and risk-based under their compliance strategy. Knowing which agents need to be turned off, or configured differently, is work that needs to happen before those agents are active, not after.

Matt Birtwistle observed that Microsoft has recently introduced new agent management technology through Agent 365, and that the industry is only beginning to work through the implications for regulated environments. Governance mechanisms for understanding how agents have been deployed, what they have done, and how to audit their activity are early-stage. For life science manufacturers leaning into this technology, those questions need to be part of the conversation now.

The Regulatory Framework That Will Shape AI in Manufacturing

For life science manufacturers trying to make sense of where AI fits in their operations, the EU’s Annex 22 is the most important regulatory development to understand right now. It is a new addendum to the EU GMP guidelines specifically addressing the use of artificial intelligence in GMP-regulated manufacturing environments. Expected to take effect in 2026, it represents the first formal regulatory framework for AI in GMP processes, and it is broadly expected to inform harmonised guidance from the FDA and other regulatory bodies in the years that follow. For manufacturers operating under EU GMP today, and for those watching where global regulation is heading, it sets the terms of the conversation.

Annex 22 introduces a framework that distinguishes between three categories of AI, each with different implications for how manufacturers can and cannot use it.

The first is deterministic software, which produces an exact, predictable output based on defined logic. Bill Burke notes that this is essentially how ERP and QMS software has always worked. The regulation formalises something the industry has been doing for decades, which means this category carries no new compliance burden.

The second is LLM-based and neural network AI, the kind of generative and predictive AI that has attracted the most attention in recent years. Here the guidance is unambiguous. Annex 11, which governs computerised systems in regulated environments and which Annex 22 builds on, is clear that this category of AI cannot perform GxP processes, including quality and operational workloads. A human must remain genuinely in the loop, not nominally so. For manufacturers excited about deploying AI broadly across their operations, this is the boundary that requires the most careful attention.

The third category sits between the other two: pattern recognition-based AI, sometimes called traditional machine learning. This is software trained to identify patterns in data and draw conclusions from them. There may be valid applications for this category in GxP contexts, but they come with an added requirement for validation and ongoing retesting to confirm the model has not drifted from accurate outputs over time. Whether and how this gets adopted against GxP workloads is a question the industry will continue to work through.

Bill’s framing of the overall challenge is a practical one: the goal is to be precise about where AI is applied, to capture the genuine efficiency gains available on non-GMP workloads, and to apply careful, documented governance around anything that touches GMP processes.

Where The Efficiency Opportunity is Clearest

It is worth being direct: the governance conversation is not a reason to slow down AI adoption. Both Bill and Matt were clear that the majority of workloads in a typical life science manufacturer are not GMP workloads. Finance, procurement, demand planning, project management, document collaboration, and several other areas are available for AI-enabled efficiency gains today, with no GxP risk involved.

Matt drew on his background as an accountant to make the point concrete. Finance teams that want to operate as genuine business advisors rather than transaction processors can start using AI for reconciliations, scenario modelling, and financial close processes right now. The tools are available. The gains are real. And staying out of GxP territory in those areas is not a constraint, it is simply accurate scoping.

Bill’s view is that organisations willing to approach this thoughtfully can realistically double the capacity of their non-GMP workloads without adding headcount. That is a significant productivity gain for companies that take the time to map their workloads correctly and apply the right tools to the right processes.

The Right Question to be Asking

For organisations evaluating or implementing technology for their manufacturing and quality operations, the most useful question to ask of any platform is not only what AI capabilities it offers. The equally important question is what controls it provides to govern which of those capabilities are active in which processes, and what visibility it gives you into that. The companies building that governance foundation now will be better positioned as the technology continues to evolve. And they will have fewer surprises when audits and regulatory reviews catch up with the pace of change.

For more information, please contact
Matt Birtwistle
Manufacturing Industry Lead
Matt.Birtwistle@nexergroup.com